Optimization of electromagnetic devices: circuit models, neural networks and gradient methods in concert

Abstract

Optimization in designing electromagnetic products is now increasingly better understood. As opposed to classical models of magnetic circuits, today, gradient techniques for mathematical optimization have been proposed and are used. These techniques, while being expensive, are exact. More recently, artificial neural networks have been suggested, but they work best only if the data set of parameter-set, performance pairs for training the network is close to the optimal solution that we seek. In this paper, it is shown how all three methods may be used in concert to increase efficiency. The circuit model is used to generate an approximate inverse solution. Then direct finite element solutions are used to generate the required training set and this is used with the neural network to get a better solution. This solution is finally used as a starting point for the gradient optimization scheme which converges quickly because the starting point is close to the actual solution.